27 / 84
tags, but never saw much acceptance
beyond that.
After a lull that lasted a couple of
decades, the topic regained some
academic status in the late 1990s and
early 2000s. By that time, theupcoming
primacy of energy consumption was
evident, and research started into
ways that commercial circuit designers
could reduce energy consumption.
Sub-threshold design techniques were
among those ideas.
The founders of Ambiq were part of
that academic revival, working at the
University of Michigan to develop the
technology more thoroughly. That
effort was spun out so that it could
be fully commercialized. Ambiq is the
only company utilizing sub-threshold
design as a primary approach to
reducing energy consumption.
It would be obvious to ask why, if
this technology was developed in the
70s, it never caught on. One might
even suspect that some flaw might
have been uncovered that kept sub-
threshold out of the mainstream. It
begs the question, “If this is so easy,
why isn’t everyone doing it?”
The answer to that question is,
“Because it’s not so easy.” There is
no fatal flaw, but the transition from
super-threshold techniques has not
been trivial. Ambiq’s founding team
started their work at Michigan in
2004 and worked until 2010 to make
the technology usable on a broad,
commercial scale.
One might also ask what’s changed
since the 70s, when the first
commercial sub-threshold devices
were created. The difference is
scale: Designs of the past used a few
critical sub-threshold transistors – on
the order of 10. At that level, each
transistor can be optimized by hand.
By contrast, Ambiq creates entire
chips that primarily use sub-threshold
transistors. That makes hand-crafting
completely impractical. Designing
millions of such transistors is possible
only by using standard design tools
and flows – preferably the same as
those that have been used for super-
threshold design. This is the work
that Ambiq has done to commercialize
sub-threshold circuits.
The challenges of
modern sub-threshold
Adapting the standard super-threshold
flows and infrastructure for sub-
threshold design presents numerous
detailed challenges. These start with
the very transistors themselves.
1. Poor transistor models
The transistor model forms the basis
of everything in an integrated circuit
design. All of the simulations, all of the
abstractions and automation, the very
process of design closure: they all rely
on an accurate transistor model. Most
transistor modeling has focused on
the “on” characteristics of the device,
with little attention given to “off.” The
entire region between 0 V and Vth
typically does not get modelled as
accurately, and so existing models are
inadequate for sub-threshold design,
as shown in Figure 2.
2. Logic swings and noise
The output response of a transistor
in the sub-threshold regime is subtle;
detecting it requires great sensitivity.
Currents change exponentially in
response to changing voltages, but
they’re exceedingly small currents.
In addition, the ratio of “on” to “off”
current is on the order of 1000,
orders of magnitude less than what
super-threshold designs experience
(see Figure 3). As can be expected,
external noise can much more easily
interfere with clean operation.
3. Sensitivity to operating
conditions
Sub-threshold designs are also far
more susceptible to process and
environmental variation than are
super-threshold designs. For example,
the current in a slow process corner
can be 10-100 times less than that for
a nominal process. Given that the on/
off current ratio (above) is only on the
order of a thousand, this cannot be
ignored.
Variations in temperature provide a
good example of how environmental
conditions create a challenge for the
designer. Vth depends on temperature,
Figure 1 - Dynamic current
dominates with higher operating
voltage
Figure 2 - Transistors haven't been
well modeled below threshold
Figure 3 - The on/off current ratio
is orders of magnitude smaller in
the sub-threshold regime